Episode Transcript
[00:00:05] Welcome to Prompting Curiosity, a podcast for the AI curious. No coding background required. I'm your host, Dr. Shantae Cofield, also known as the Maestro, and I created this show to explore what these AI tools actually are. Really, though, are the files in the computer, how to use them, and what they might mean for how we think, work, create, and move through life. Whether you're skeptical, intrigued, or already experimenting, you're in the right place. All that I ask is that you stay curious. All right, let's get into it.
[00:00:38] Hello. Hello, my curious people, and welcome to episode 53 of Prompting Curiosity. So I actually, as I'm saying this, I'm realizing 52 weeks in a year, we are officially past one year of content.
[00:00:52] Thank you for being here. I know that I celebrated episode 50, which I'm absolutely going to celebrate. Uh, but one of the things I talk to people about when it comes to making a podcast, I'm like, give yourself one year, commit to a year, and then you can see what you want to do with it after.
[00:01:06] Uh, I obviously have no, um, I have no plans to, like, stop this. It wasn't like, oh, give it a year and see. Like, I like yapping. So here I am. But we are past a year of content, and I'm super grateful that you are here. So if you don't know, I am your most grateful host, the Maestro. And today we're talking about something that has been around for, um, quite a bit. A little bit. Um, but I have yet to actually try it until last week, and that is deep research. So this is literally. And by this, I mean deep research is literally a way to have the robots do your homework for you.
[00:01:45] And that is what I'm going to call a 100% Maestro approved use of AI. So, you know, I don't like having AI speak for me or write for me or think for me, but I am definitely cool with having AI search for me and also code for me. Uh, and that's what we're diving into today. So we're gonna get right into it. Definition coming up. First, Deep Research. What is it? Deep Research is an AI mode that autonomously searches, reads and synthesizes dozens to get this, hundreds of web sources over several minutes to produce a structured, cited report on a topic, rather than instantly answering from what the model already knows. Right. Every LLM has its own version. Google, OpenAI, Anthropic. They all have their own version. They introduced them roughly 20, 24, 25. So first off, I believe was Google that launched in 2020. Their Their research mode, Deep Research mode launched, um, in December 2024. OpenAI followed in February 2025. Anthropic came after that, um, in April of 2025. They call theirs Research as opposed to Deep research, um, perplexity 2025 and then Xai, uh, aka Grok, that uh, in 2025. So again, deep Research is a mode, right? You just, you select it, you turn it on and it is different than a chatbot reply. The normal things that we get, right, based on your prompt, whatever you put in there and you ask it for the model is not just going to answer from what it already knows. It will go and search the web. Sorry about that whistle there, got a gap in my teeth. It will go and search the web, it will read actual sources, it will cite them, and it will generate a written report report for you. This is really cool.
[00:03:26] Uh, when you run this, it typically runs from like a few minutes, like five minutes to like 30 plus minutes depending on what the task is. And during that time you can 100, you just walk away. It will do its thing and then the final output will be this structured report, usually like a few pages long. And it includes citations like that's, that's actually really cool. So how it works without getting like techy at all.
[00:03:50] So what we have at the front end is a leap. You put your prompt in and then a lead agent is going to read that question, quote, unquote, read the question. Uh, and it writes a research plan. So I'm going to refer you back to last week's episode as episode 52, AI agents explained for non coders. Uh, if you want to understand more about agents, AI agents. But again, it is a, uh, it's a computer program right within the LLM wrapped around the LLM.
[00:04:16] But this agent is going to go and um, read the question and make a plan. From there it's going to break the plan into chunks and it will hand each chunk to what's called a sub agent. So just another program that's going to be doing that specific part.
[00:04:31] And that sub agent will search its specific slice of the topic.
[00:04:36] Because the agents are searching all of these sources, Remember like dozens to hundreds of sources.
[00:04:42] This process will take minutes instead of seconds. Right. You just ask Claude or ChatGPT a question, regular question. It's not going and searching something, it's just giving you an answer based on what it knows. Yes, every now and then it will like go and like look at things from the, from the Internet, but it's not searching dozens to hundreds of sources. And it's not synthesizing a report from that, right? So we are currently at the step where the all of these little agents are, are searching the interwebs, looking at sources.
[00:05:08] Those agents then come back with their findings and they bring back the sources that they pulled those findings from.
[00:05:14] A synthesizer then stitches all of that into one report and there's a citation pass that makes sure that the uh, the claims that are stated, they point back to where they came from, right to the sources. So we see it's quite a process that goes on, which means it's going to be more token heavy, more cost intensive. Um, but it's pretty fucking cool because it produces a report like that's pretty dope. So as per always per always better input equals better output. So you know, a good request would be, would include things like what do you want, what do you want to include, what sources prioritize, you know, what format, who is it for, as much information, as much specifics as possible. A bad uh, request would be like tell me about AI note taking tools. Like you could do that, but not the best use case, right? It's you're not guiding this thing as much as as you can or as much as you should. So the best input, right, would be to ask AI to, to write it for you, right? To ask it to write the prompt for you. But an example of a better um, a better prompt would be compare. Three AI note taking tools for solo creators include pricing, privacy, transcription, quality, prioritize, vendor docs, and rep. Reputable reviews. Vendor docs meaning it's like from the actual company that's making the thing. It's not just like you know, an affiliate um, article, right? And then you could say give me a table. You can say how you want this information, uh, presented, right?
[00:06:41] Bonus little tip here. Some of these tools, depending on what LLM you're using, they will let you restrict the search to certain sites, um, or point it at your own files. So it's not just grabbing anything that's on the open web. This is very, feels very akin to um, feels very akin to Google Notebook lm. I will link that episode as well. Um, if you're not using Google Notebook lm, you don't know what it is. Definitely go check out that episode because it is a very, very cool tool.
[00:07:10] Um, and so I'm gonna leave with that. Uh, next up, some example, use cases because like many things with AI deep research, to me at least feels a bit like here's a solution, go find A problem, particularly for folks like us, perhaps, that don't. And I'm assuming that, you know, you're similar to me. Um, but if you have a job that doesn't really involve researching and comparing things, you're just like, what the fuck am I using this for? Right?
[00:07:35] So my friend K, he works in finance and he has been using deep research, like from the beginning.
[00:07:42] And that's largely because so much of the work that he's involved with or so much of the work that he does involves researching things like, you know, company information and numbers and, you know, publicly available stats and things like that.
[00:07:54] All right, so this is a great use case. That's a great use case. And he's noted time and time again how much time it has saved him. But for folks like us, I think there's less obvious use cases. And admittedly I only used research mode so that I would have some experience with it for this episode because, um, I didn't have like a set, you know, super burning use case. But, um, I did try. I did, I did, you know, use it for that. And I'll go into, in the end of the episode I'll talk about how I used it. But, um, in this particular section right now, I just wanted to give you some examples of things that you could use it for. So you can use it to vet a tool or software before you buy it. Uh, you can look at comparison with pricing and privacy.
[00:08:38] Uh, you can use it to, um, research what competitors charge and how they position.
[00:08:45] You could use it for pre call prep, brief a person briefing on a person or a brand or a company before sales, call before a podcast. Um, you can, you know, this could be, it doesn't have to be like right before the podcast. It's going to be part of your actual preparation. Um, you can use it to figure out a new, a new platform or channel. And I like the suggestion, right. I obviously asked Claude for some, some examples and I was like, oh, okay. I picked the ones that I actually liked and might, we might use it for, um, but figuring out a new platform or channel, right? So example, is substack worth it for someone like me or what's actually working on YouTube in my niche. And just as a starting point, you know, to give you a report on that.
[00:09:19] Um, you could use it to find the questions that people are actually asking about your niche moving outside of work. You could use it for big purchases like a car or stroller mattress. You know, the thing that you're overthinking, but because it like costs a lot of money. Um, you could use it to compare health insurance plans. Right. The through line here being that we're not going to use it necessarily for like some 40 page research analysis, um, as people in like, you know, real research, heavy industries may use. It's that we're going to use it to produce a report that saves us hours of, you know, tab tab hopping. Go from this tab to this tab to this tab to try and you know, compare things and it'll create, you know, a brief that we can actually act from. Like that's, that's pretty dope. So again, all of the major players, major AI players have deep research or research mode. Um, you, if you, whatever one you're using, it has it. Um, um, they do tend to, what's the word I want to use? Excel, I don't know in, in different areas.
[00:10:20] So ChatGPT, um, this is apparently the most comprehensive one. The longest reports are produced, the most sources and thus it's also gonna be the slowest. Gemini fast and broad because it's plugged directly into Google Search and apparently it exports straight to Google Docs and it has an actual free tier. Claude typically uses fewer sources, but it's strong at reasoning through sources that contradict each other and it does produce CLE writing and then last grock that it pulls live from X. Who wants that? I don't know, but I'm just letting you know. Um, as a heads up, using research mode for any model will eat way more of your usage credits than normal chat.
[00:11:05] Um, I'm not even sure fully how that works with, I'm not sure how that works with ChatGPT because as I'm saying this I'm like, I actually didn't look into this. You may need to be on the higher tier, like the 200amonth tier, um, to use it for chat GPT, you do not need to be on the higher tier. For Claude, I am on the 20amonth tier and, and I have it. But obviously Claude has usage limits and it definitely uses more usage uh, than just regular chat. Makes sense. Right? It is using agents, right? Or it does have agentic properties as well. Right.
[00:11:39] Um, um, so just use it wisely.
[00:11:41] So how to actually use this thing depending on the tool that you're using, meaning depending on the AI tool. So whether you're on chat, uh, you know, Chat GPT or, or Claude, there will be a mode selector somewhere. Usually it's like in or near the message box or like the prompt field. Usually um, like the behind the plus sign. And then from there you're just gonna select research or deep research, whatever it's called, depending on the tool. And then you'll prompt it as normal. As per always. I recommend having AI create the actual prompt for you. Right. And do not be surprised if it's very comprehensive. So when I did this, uh, to just try it out, I had Claude come up with the prompt and I was like wow, this is, I told m him. I was like, this is, you know, you see what I think I would use it for.
[00:12:20] Um, can you generate a prompt that I can, can paste? And it gave me a really good one and I was like, oh wow, this is like very comprehensive. So uh, as per always, my, my recommendation is to have the, the robot teach you how to use the robot and tell you how to use the robot and come up with the language so that you can use the robot. So um, some models, Claude didn't, but some models will show you a research plan before it starts. Um, like we talk about this with Claude code like before it's in planning mode and it shows you. Okay, I'm going to go pull from these sites. I'm going to do this thing that did not happen in when I, when I did research mode. Um, perhaps other models use, other AI models use it. Um, but just be aware that is a thing and that it is helpful in my opinion. Like, okay, cool, I see the plan and then you can uh, accept the plan or you can modify it and then, and then go from there.
[00:13:11] If you are able to choose, I suggest choosing the strongest reasoning model within that specific LLM that you're using. Right. So the strongest reasoning model within Claude, the strongest reasoning model within ChatGPT. And so sometimes you cannot choose, but if you can choose, I would suggest using the strongest reasoning model. Why? Because it has to plan a multi step job. It has to judge with source which sources are worth trusting and it has to synthesize this, you know, possibly conflicting points into something coherent. So you want some, you want a model that has the best reasoning skills and capacity for that. So remember, uh, clearly that's also going to make it even more token expensive or token intensive. Um, so for this I used Opus 4.8.
[00:13:57] Yes, there is a tier higher. I did an episode on it, it's called Fable. But that felt like a little bit like overkill for this. But go ahead, knock yourself out. Um, again it will use more tokens. Um, but I went with Opus 4.8 and M, my daily driver continues to be Sonnet 5 so, you know, the thing I want you to take away is use the, um, strongest reasoning model. Right. Again, the model will then run anywhere from a few minutes up to like 30 plus minutes. When I did it, it was like, I don't know, six or seven minutes. Um, it just depends on the task assigned. You don't need to be there sitting there watching it, answering any questions. You just, it just does its thing and then you come back and it will have produced a comprehensive report. That's pretty dope.
[00:14:36] A thing to remember is that it's still an LLM, right? So it can still hallucinate, AKA make things up. Uh, it can still cite a bad source and do so very confidently. It can. And it will miss paywalled or gated content that it can't access. That may be very, very important. So, you know, as of her, always check the work, check the robots work, read the citations, verify the things. Right.
[00:15:03] Um, yeah, that is, that is. I'm like, went through that so fast. I'm like, okay, yeah, that is it. Um, because the segue is into how I used to AI this week, which, if you're new here, welcome. Uh, each episode I share a quick example of how I used AI this week or that week. This week I used Claude in research mode. Uh, and mainly I was using it to just test it out. So I wasn't just, you know, spewing things at you from theory here, but I used it to identify and highlight content gaps in the AI space.
[00:15:35] Uh, and from there, suggest episode topics for this podcast. So my review, the process, it was good, right? The report was very good. It did its job. It cited solid sources, it organized things very cleanly.
[00:15:48] It did a great job. The asterisk here is that I didn't love the results, not because they were bad, but because I don't want to make episodes about the topics that had surfaced, which is 100, you, uh, know, 100% a me thing. I was like, I don't fucking talk about that. So that's, that's fine. I don't have, I don't have to take what the room with the report says. It's just like, okay, those are gaps.
[00:16:08] You know, fill them if you want, don't fill them if you don't want to. And I will not be filling them. But I do, however, think that overall, this is very much a net positive tool. And when a use case arises, I will gladly have the robots do my homework for me. All right, that is all for today. Hopefully you found this episode helpful. If you did consider sharing it with somebody you know who's curious about AI, and maybe they could use a little heads up about Deep Research mode. Don't forget I have a companion newsletter and blog called the Curious Companion and that drops every Thursday. And it's basically my basically I mean exactly the podcast episode in text format. So if you prefer to read or you just want a written record of things, join the newsletter fam. Go check out the vlogs. You can head to prompting Curiosity.com forward slash newsletter or forward slash blog. Or just check out the link in the show notes. Easy peek. Easy. I gotcha. As always, my friends, endlessly, endlessly, endlessly appreciative for every single one of you. Until we chat again next Thursday, stay curious.
